================
@@ -3282,6 +3292,36 @@ bool ARMExpandPseudo::ExpandMI(MachineBasicBlock &MBB,
MI.eraseFromParent();
return true;
}
+ case ARM::WIN_RECOVER_FP: {
+ MachineBasicBlock &MBB = *MI.getParent();
+ MachineBasicBlock::iterator MBBI = MI.getIterator();
+ MachineFunction &MF = *MI.getMF();
+ DebugLoc dl = MI.getDebugLoc();
+ Register DstReg = MI.getOperand(0).getReg();
+ const ARMBaseRegisterInfo &RI = TII->getRegisterInfo();
+ Register BaseReg = RI.getBaseRegister();
+ Register LocalAddrReg = RI.getLocalAddressRegister(MF);
+ if (AFI->isBasePtrSpilled()) {
+ // If the handler itself uses R6, we must restore the parent's R6.
----------------
trungnt2910 wrote:
I was worried that `r6` might get clobbered when the funclet has its own base
pointer. However:
1. `__except` blocks are just a special label in the parent function.
2. `__finally` blocks are separate functions. They don't access variables via
`r6` but through a live-in `r1`. They also don't set up the pointer - it's
either the parent function or the trampoline that calls the `__finally` block.
3. The trampoline before `__finally` block does nothing but pass `r6` into `r1`
for `__finally` blocks to consume. They can't define aligned variables or VLAs.
4. `__except` filter functions. These are tricky because they allow arbitrary
expressions, which may include aligned variables and VLAs via a
`[[msvc::forceinline]]` lambda. However:
- Realigned variables can still be accessed via `sp`, if no VLAs are
involved.
- VLAs are already not supported by `clang` when in an SEH context:
https://godbolt.org/z/hTjbaP8bx. They are also forbidden by MSVC.
There is also code that prevents clobbering `r6` due to it affecting other
funclets like the case in `3.`:
https://github.com/llvm/llvm-project/blob/9889b965d18c37601fbe225adc0041c75ac3be9c/llvm/lib/Target/ARM/ARMFrameLowering.cpp#L1384
I added `CHECK` lines for `4.` and asserted that the aligned variable is
accessed through the stack pointer `sp` and that `r6` remains unclobbered.
https://github.com/llvm/llvm-project/pull/184953
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
